band_snr — MOTIONMAG temporal op

Data kinds: videotable

Call: import motionmag; motionmag.band_snr(video, f_lo, f_hi, fps) -> 'dict' (or opsmotionmag.get("band_snr"))

Usage

Measure what a clip's temporal band contains, and what it costs -> `dict`.

Every quantity is a measured mean-square power obtained from the per-pixel

temporal DFT (Parseval-normalised so that the bins of one pixel sum to that

pixel's mean square), averaged over pixels:

• `static_power` — the DC bin. The scene that is simply *there*.

• `band_power — the bins inside [f_lo, f_hi]`. Coherent motion plus

whatever noise happens to fall in the band.

• `out_of_band_power / out_of_band_bins` — everything else except DC.

With broadband sensor noise this is the noise floor, and

`noise_power_per_bin` is its per-bin density.

• `noise_in_band = noise_power_per_bin * band_bins` — how much of

`band_power` is expected to be noise.

• `motion_power = max(band_power - noise_in_band, 0)` and

`motion_snr_db = 10*log10(motion_power / noise_in_band)`.

• `image_snr_db = `10*log10(static_power / (band_power +

out_of_band_power))`` — the static scene against everything that flickers.

**The two SNRs answer different questions and magnification moves only one

of them.** Scaling the in-band phase by `alpha` scales the in-band motion

*and* the in-band noise by the same factor, so the true motion SNR cannot

improve: magnification never makes a measurement more certain than the

recording was. What does change is `image_snr_db`, because the temporal

fluctuation of the output frames grows like `alpha^2` while the static

scene does not.

A caveat that matters when this is run on an already-magnified clip.

`motion_snr_db` here divides the in-band signal by a noise floor estimated

from the *out-of-band* bins, and magnification does not touch those. Applied

to a magnified video it therefore credits `alpha^2` more in-band power

against an unchanged noise estimate and reports an improvement that did not

occur — measured, `+6.86 dB at alpha = 2` on a clip whose true motion

SNR cannot have moved. :func:motion_magnify knows the gain and returns the

corrected figure as `motion_snr_out_db`; use that one, not

`result["snr_out"]["motion_snr_db"]`.

`snr_clamped is True when a reported dB hit the [-100, +100]` window

(a noiseless synthetic has zero out-of-band power, which is a division by

zero rather than an infinite SNR).

Detailed usage guide

motion_magnification family guide

References (sample data, literature)

• Sample-data catalog (download URLs / licences) — 2-D uses skimage.data (BSD/public domain) plus synthetic images; 3-D lists download URLs for real data sources (Stanford, PDS, …).

• Operator provenance and references — the sources of the research/methods this op family came from.

• The canonical algorithm (author, year) and its uses are named in the family usage guide above.

Runnable examples (verified samples that actually call this op)

motion_magnificationpy -3.11 examples/motion_magnification.py

Ops the type connects to (they accept table as input)

complex_steerable_reconstruct

Same category (temporal)

temporal_bandpass · temporal_band_power


*Provenance: motionmag.py — MOTIONMAG operator registry. This per-op note is generated by tools/opdocs.py md (do not hand-edit).*

© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.